{"id":"412c3953-3117-43ab-ab4c-66fbb9ef86a3","arxiv_id":"2411.19032","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A review of machine learning applied to spectrum sharing, covering sensing, allocation, access, handoff, beamforming, and security, with summary tables of the literature.","lead":"This paper surveys how machine learning methods are being used to share wireless spectrum among competing users and technologies. It maps existing work onto spectrum sensing, allocation, access, and handoff, and lists open research questions.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'complete overview' claim is unsupported because the survey documents no systematic literature search or inclusion criteria; a bibliometric recall check against 2020-2024 databases would settle it.","rationale":"The reader's weakest assumption identifies exactly the load-bearing vulnerability: the survey's central claim is completeness, yet no systematic literature search or inclusion criteria are described. My stress-test pass found no additional concern that changes the verdict. The survey's internal organization is coherent, the ML background section is mathematically standard, and the per-chapter tables are useful summaries; I found no internal inconsistency that would make the survey misleading beyond the coverage claim. The 'complete overview' statement in the abstract and the checkmarks in Tables 1.1-1.2 are empirical claims that cannot be verified from the manuscript as submitted. The proposed bibliometric audit would test that claim directly: if a substantial fraction of recent topic-relevant papers are absent, the survey should be repositioned as a curated rather than complete overview. Because the reader already assigned CONDITIONAL with moderate confidence on this basis, my recommendation is UNCHANGED rather than a movement to ACCEPT or REJECT. The secondary concern about table-summary accuracy is worth a spot-check but does not by itself change the verdict, since a single mischaracterized row would not invalidate the survey's organizational value.","tokens_in":47750,"tokens_out":2173,"duration_ms":22818,"concrete_test":"Perform an independent bibliometric recall audit. Construct a reproducible query for 2020-2024 in arXiv, IEEE Xplore, and Scopus, e.g., (('machine learning' OR 'deep learning' OR 'reinforcement learning') AND 'spectrum sharing'), plus per-topic variants covering sensing, allocation, access, handoff, beamforming, and security. Take the top 50-100 results per topic by relevance or citation count and check each against the survey's reference list and summary tables. Report recall per topic. If recall is below a pre-specified level (e.g., 80%) or any topic systematically misses recent work, the 'complete overview' claim fails. As a secondary check, randomly sample 20 summarized rows from Tables 4.1-4.3 and 5.1-5.3 and compare each one-line description with the cited paper's abstract or stated findings; flag any factual mismatch.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the abstract's 'complete overview of the state-of-the-art of machine learning for spectrum sharing' and Section 1.1's assertion that this survey differs from prior surveys by covering ML fundamentals, all four sharing mechanisms, beamforming, and security. For that claim to hold, the set of surveyed papers must accurately represent the actual literature. The paper never describes a systematic search: no databases, query strings, inclusion/exclusion criteria, time window, or screening process is given. Section 1.1 only enumerates prior surveys [24]-[40] and states their gaps, while Tables 1.1-1.2 mark 'Our work' with checks for every mechanism; that is an assertion, not a demonstrated coverage result. Figure 1.3's keyword density map is described as showing the most relevant topics, with keywords mentioned in at least five references, but no protocol explains how the underlying reference set was assembled. A reader therefore cannot audit whether important recent work is missing, e.g., transformer-based spectrum sensing, generative-AI spectrum allocation, digital-twin-assisted sharing, or new deep RL algorithms from 2023-2024. The survey's value as a reference map is conditional on this coverage claim. A secondary concern is that the one-sentence summaries in tables such as Tables 4.1-4.3 and 5.1-5.3 are second-hand interpretations; if any mischaracterize the cited work, the comparative organization inherits the error. Neither concern implies internal inconsistency in the ML exposition, but both bear directly on the paper's advertised completeness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a survey of machine learning (ML) methods applied to spectrum sharing. It opens with an introduction to supervised learning, unsupervised learning, and reinforcement learning, then reviews ML-based solutions for spectrum sensing, spectrum allocation, spectrum access, spectrum handoff, beamforming, and security. The paper compares its scope with earlier surveys and presents summary tables of the surveyed works, a keyword-density map, and a discussion of open challenges. The abstract and Section 1.1 claim that the paper provides a complete overview of the state of the art in ML for spectrum sharing.","tokens_in":48010,"tokens_out":4905,"duration_ms":81628,"significance":"If the coverage is representative, the survey would be a useful reference map for researchers entering this area. The paper has several strengths: it includes a self-contained mathematical tutorial on the main ML families, it covers all four spectrum-sharing mechanisms plus beamforming and security, it provides comparison tables that organize the literature, and it closes with a concrete list of open problems. These are valuable features for a survey in a journal-style monograph series. However, the central claim of completeness is not supported by an auditable literature-selection procedure, and the ML tutorial contains several mathematical errors in equations that are later used as conceptual foundations for the surveyed applications. These issues do not require redoing the survey, but they do require revision before the paper can serve as a reliable reference.","major_comments":[{"comment":"The abstract and Section 1.1 claim a 'complete overview of the state-of-the-art of machine learning for spectrum sharing,' but the paper never describes how the reference set was assembled. No databases, query strings, time window, inclusion/exclusion criteria, or screening process are reported, and Fig. 1.3 refers to a 'considered database' without defining it. As a result, a reader cannot audit whether the claimed coverage is complete or biased. I recommend adding a reproducible search protocol and, ideally, a bibliometric recall check against 2020-2024 literature, or softening the completeness claim to reflect the actual selection procedure.","section":"Section 1.1, Tables 1.1-1.2, Fig. 1.3"},{"comment":"The between-class covariance matrix is written as S_B = (m1 - m1)(m1 - m1)^T, which is identically zero. It should be (m1 - m0)(m1 - m0)^T. As written, the Rayleigh quotient in Eq. (2.7) is degenerate and the Fisher discriminant analysis tutorial is incorrect. This is a load-bearing error because the paper explicitly advertises a mathematical description of ML methods as one of its contributions.","section":"Section 2.1.3, Eq. (2.5)"},{"comment":"The Q-learning update in Eq. (2.57) uses 'arg max_a q_pi(s_{t+1}, a)' where the update target should involve 'max_a q_pi(s_{t+1}, a)', i.e., the value, not the maximizing action. Similarly, the deep Q-learning update in Eq. (2.61) uses the next action a_{t+1} in the target term, which is the SARSA target rather than the DQN target; the correct target is r_{t+1} + gamma max_a q(s_{t+1}, a; w_t). Since Q-learning and DQN are the central algorithmic themes of Chapters 4 and 5, these errors weaken the tutorial foundation of the survey.","section":"Sections 2.3.2 and 2.3.3, Eqs. (2.57) and (2.61)"},{"comment":"The table summarizes references [151]-[154] as 'Satellite IoT' and as decreasing DQN instability via experience replay and target networks. The body text, however, describes [151] as a NOMA channel-assignment problem, [152] as a vehicle-to-vehicle network, and only [153] and [154] as satellite systems. This concrete mismatch shows that the one-sentence table summaries can mischaracterize the cited papers. I recommend checking every table entry against its source, because the comparative value of the survey depends on the fidelity of these second-hand summaries.","section":"Table 4.2, row [151]-[154]"}],"minor_comments":[{"comment":"The chapter heading reads 'Introduction to Machine Leaning' and should read 'Machine Learning'.","section":"Chapter 2 title"},{"comment":"The text says the thresholds are for 'class H0 and H0 respectively'; this should be H0 and H1. It would also help to verify the direction of the inequalities, since the sentence 'if channel is free' and 'if channel is busy' can be read in either order depending on the chosen encoding.","section":"Section 3.2, Eq. (3.2)"},{"comment":"The entry ends with '6G systes' and should read '6G systems'.","section":"Table 1.2, row [38]"},{"comment":"The comment column reads 'Proposed a The DDQN algorithm' and should be corrected to 'Proposed a DDQN algorithm'.","section":"Table 5.2, row [120]"},{"comment":"Several entries contain the typo 'handoof' instead of 'handoff', and Section 6.1 spells 'Nave Bayes' instead of 'Naive Bayes'.","section":"Table 6.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a suitable fit for this monograph series and is likely to be useful after revision. My main reservation is the gap between the advertised 'complete overview' and the undocumented reference-selection process; I would ask for a reproducible search protocol or a softened claim. The tutorial errors in Eqs. (2.5), (2.57), and (2.61) are local and fixable, so I do not view them as grounds for rejection, but they should be corrected before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis is a competent, clearly organized survey of ML for spectrum sharing, and it is probably useful as an entry point for a grad student or engineer new to the area. What it is not, despite the abstract's promise, is a demonstrably complete overview of the state of the art.\n\nWhat is actually good: the paper does more than the four-part taxonomy (sensing, allocation, access, handoff). It adds a self-contained ML primer with mathematical formulations that are often skipped (empirical risk, SGD, neural network layers, Q-learning), and it extends coverage to beamforming and security, which prior surveys like [36] did not include. The summary tables in Chapters 4 and 5 are genuinely useful for comparing RL-based allocation and access methods. The authors also engaged honestly with 17 prior surveys, explaining where each falls short.\n\nWhere it is soft: the completeness claim rests on an undocumented reference selection. There is no description of databases, query strings, inclusion/exclusion criteria, or a time window. The keyword density map in Fig. 1.3 is described as showing keywords occurring in at least five references, but there is no protocol for how the reference set was assembled. So a reader cannot audit whether 2023-2024 work (transformer-based sensing, generative AI allocation, digital-twin sharing) was missed. That is the load-bearing weakness, and it is fixable by moderating the language or adding a search methodology. The one-sentence table summaries are second-hand interpretations; if any mischaracterize a cited work, the comparison inherits the error. There are also several typos (Eq. (3.2) lists H0 and H0 instead of H0 and H1; Eq. (2.5) uses m1 in both terms). These are minor but suggest a final proofread pass.\n\nThe ML exposition is standard and accurate; there are no derivations to validate or fitted parameters, so the circularity burden is zero.\n\nWho this is for: newcomers to ML-based spectrum sharing who want a structured map and quick comparison tables. It will not settle any open research problem, and it should not be cited as evidence of completeness. I would still send it to peer review because the tables and cross-survey comparison give it enough substance that a referee can meaningfully check coverage and recommend revisions. I would ask the authors to tone down the complete claim and either document the search or reframe it as a representative selection.\n\nRecommendation: engage with it, but treat the completeness claim as a hypothesis the authors need to justify, not a fact.","headline":"Competent, well-organized survey whose 'complete overview' claim outruns the undocumented reference selection; useful as an entry point, not as a definitive audit.","tokens_in":48561,"tokens_out":2541,"would_cite":true,"duration_ms":23357,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This survey argues that the state of the art of machine learning for spectrum sharing can be organized into four mechanisms—sensing, allocation, access, handoff—plus beamforming and security, and that the work in each area clusters into…","keywords":["spectrum sharing","machine learning","spectrum sensing","spectrum allocation","spectrum access","spectrum handoff","reinforcement learning","dynamic spectrum access"],"falsifier":"A systematic literature review of ML for spectrum sharing using explicit inclusion criteria from the same databases and time window would settle the claim: if a substantial fraction of qualifying papers fall outside the paper's four mechanisms, beamforming, security, or its ML taxonomy, or if the table summaries mischaracterize several cited works, the central organizational claim is falsified.","tokens_in":47598,"feed_emoji":"📡","tokens_out":4070,"duration_ms":36163,"temperature":0.7,"pith_summary":"The paper tries to establish a complete map of how machine learning is being used to share wireless spectrum in 5G and 6G networks. It organizes the field into four mechanisms—sensing, allocation, access, and handoff—and adds beamforming and security, then sorts the literature by learning paradigm. A sympathetic reader would take the central claim to be that this taxonomy captures the state of the art and reveals which combinations of ML method and spectrum-sharing task are mature and which are open. If right, it gives researchers a structured way to place new results and spot gaps.","feed_headline":"Survey maps ML onto all four spectrum-sharing stages","feed_subtitle":"Sensing, allocation, access, handoff, beamforming and security are filed by learning family, with gaps flagged.","key_machinery":"The organizing object is the four-mechanism pipeline—spectrum sensing, spectrum allocation, spectrum access, and spectrum handoff—that connects the radio environment to the user. The survey uses this pipeline together with the division of ML into supervised, unsupervised, and reinforcement learning; each chapter maps one mechanism onto the learning families, with tables summarizing the algorithm, network type, and contribution of each cited work.","core_discovery":"The discovery on offer is not a new algorithm but a comprehensive classification: recent ML-based spectrum sharing can be seen as the application of three learning families to four coordinated mechanisms. The paper provides mathematical formulations of the ML methods in the spectrum sharing context, surveys and tabulates supervised, unsupervised, semi-supervised, and reinforcement-learning solutions for sensing; Q-learning and deep Q-learning plus other RL for allocation and access; and identifies handoff, beamforming, and security as further aspects with emerging ML contributions. It claims to cover what prior surveys miss: ML fundamentals, all four sharing mechanisms, beamforming, and security in one place.","pith_inferences":["Beyond the paper: the 'complete overview' claim rests on an unsystematic reference selection, so the map is best read as representative rather than exhaustive.","Beyond the paper: because most surveyed studies use simulation with different metrics, the taxonomy says little about which ML solution performs best in deployment; a shared benchmark suite would test the field's practical progress.","Beyond the paper: the gap pattern suggests that the least explored combinations—such as unsupervised learning for spectrum access or ML for handoff—are the likely sites of the field's next experiments."],"forward_implications":["A new spectrum-sharing solution can be located in the taxonomy, making systematic comparison with neighboring methods straightforward.","The mathematical formulations reveal a common structure: classification for sensing, and Markov decision processes or reinforcement learning for allocation and access.","If the coverage is accurate, spectrum sensing is the most mature ML area, while handoff, beamforming, and security are comparatively thinner and are candidate directions for future work.","The tables offer a quick reference for which RL algorithms—Q-learning, deep Q-network, double DQN, DDPG, PPO—have been applied to which network types."],"supporting_citations":[{"why":"Supplies the narrowband and wideband sensing taxonomy that the sensing chapter organizes and extends.","marker":"[24]"},{"why":"Provides the deep-learning spectrum sensing survey whose coverage the paper positions itself against.","marker":"[25]"},{"why":"Baseline survey of reinforcement learning for spectrum allocation that the allocation chapter builds on.","marker":"[28]"},{"why":"Baseline on spectrum sharing techniques for CR, D2D, IBFD, NOMA, and LTE-U with which the paper contrasts its ML focus.","marker":"[29]"},{"why":"Prior ML-focused survey of cooperative sensing and dynamic spectrum sharing that the paper extends to all four mechanisms.","marker":"[33]"},{"why":"Closest prior classification-and-survey of ML for spectrum sharing, used as the main comparison point.","marker":"[36]"},{"why":"Prior survey of ML-based mmWave beamforming that the paper broadens to all frequency ranges.","marker":"[37]"},{"why":"Baseline survey of reinforcement learning for physical-layer security, which the paper complements with supervised and unsupervised security works.","marker":"[38]"},{"why":"Provides ML defensive strategies for spectrum sharing security that the security section draws on.","marker":"[39]"}],"fun_headline_variants":["ML survey maps all four spectrum-sharing stages","Survey: ML for sensing, allocation, access, handoff","Three ML families, four spectrum tasks, one survey","Survey unifies ML for spectrum-sharing stages"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's value depends on the selected references being a faithful picture of the whole field, since the authors describe no systematic literature search or inclusion criteria; if the selection is skewed, the map looks complete when it is not.","fun_headline_variants_meta":{"raw":{"variants":["ML survey maps all four spectrum-sharing stages","Survey: ML for sensing, allocation, access, handoff","Three ML families, four spectrum tasks, one survey","Survey unifies ML for spectrum-sharing stages"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000841,"raw_usage":{"total_tokens":3671,"prompt_tokens":961,"completion_tokens":2710,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":2649}},"tokens_in":577,"tokens_out":2710,"duration_ms":34990,"temperature":1.0,"reasoning_tokens":2649,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:35:38.526526+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A systematic literature review of ML for spectrum sharing using explicit inclusion criteria from the same databases and time window would settle the claim: if a substantial fraction of qualifying papers fall outside the paper's four mechanisms, beamforming, security, or its ML taxonomy, or if the table summaries mischaracterize several cited works, the central organizational claim is falsified.","supporting_citations":[{"cited_title":"A comprehensive survey on spectrum sensing in cognitive radio networks: Recent advances, new challenges, and future research directions,","cited_arxiv_id":null,"evidence_quote":"Supplies the narrowband and wideband sensing taxonomy that the sensing chapter organizes and extends."},{"cited_title":"A survey of dynamic spec- trum allocation based on reinforcement learning algorithms in cognitive radio networks,","cited_arxiv_id":null,"evidence_quote":"Baseline survey of reinforcement learning for spectrum allocation that the allocation chapter builds on."},{"cited_title":"Machine learning for cooper- ative spectrum sensing and sharing: A survey,","cited_arxiv_id":null,"evidence_quote":"Prior ML-focused survey of cooperative sensing and dynamic spectrum sharing that the paper extends to all four mechanisms."}],"review_version":1}